1. Field of the Invention
This invention relates to probabilistic neural networks, and more particularly relates to an adaptive probabilistic neural network that can sort digitized input pulse parameter data signal description words without the use of a priori training data.
2. Description of the Prior Art
Radar emitter pulse sorting and radar emitter identification are the primary functions of electronic support measure (ESM) and electronic counter measure (ECM) systems. There are three basic steps involved in the emitter identification process. First, the input pulse signals undergo an initial level of analysis and differentiation commonly referred to as "sorting" or "pulse-by-pulse deinterleaving". The sorting process involves analysis of the input signals to achieve an initial grouping of pulses from each emitter in the collected pulse sequence. If a high percentage of pulses are correctly sorted and grouped during the first sorting level, then only a small number of pulses will undergo a second level of deinterleaving. However, invariably many input pulse signals are not capable of being correctly sorted because the signals can not be easily differentiated by the system. The sorting system is not able to recognize the input signals because the input signals are often noisy, inaccurate and corrupt with additional or missing signal parameter components or information. The pulse groups which were not correctly sorted and grouped by the system at the first level of analysis require a second level of analysis commonly referred to as "second level deinterleaving". This second processing and sorting level requires multiple and complex sorting algorithms which occupy a great deal of computer time. Once all of the input signals have been sorted and deinterleaved by the first and second levels of analysis, they are transferred to a third stage of processing commonly referred to as emitter identification. During this stage, the sorted groups are analyzed so that the radar emitter transmitting each type of signal can be identified for ESM and ECM purposes.
In the past, various rule-based techniques were developed for sorting digitized pulse signals. One of the earlier rule-based sorting systems is commonly referred to as the histogram method. The histogram method compares each input pulse parameter signal against a group of preset signal parameters. The comparison is performed to determine if the parameters of the input pulse parameter signal can be classified within the group of preset signal parameter values. However, the histogram method may not accurately sort the incoming signal when even one parameter of the input pulse parameter signal does not match the preset signal parameter values. This makes the histogram method undesirable. The histogram method is also undesirable because the incoming signals must be input to the system at a relatively slow rate as compared to the rate that the pulse signals are transmitted by the radar emitter. Therefore, a sorting system utilizing the histogram method is not readily able to produce a real time system response to incoming radar pulse signals.
Another early rule-based sorting technique is commonly referred to as "adaptive binning." Adaptive binning compares individual parameters of the input pulse signal to preset signal parameter values. Each input pulse signal can have numerous parameter values. The adaptive binning system is relatively slow in operating because only one parameter comparison is undertaken at a time. Therefore, successive comparisons are not made until preceeding comparisons are complete.
Additionally, the adaptive binning system is very rigid, inflexible and incapable of sorting input signals having parameter value errors. For example, if an input pulse signal consists of ten parameters, and one parameter of the group of ten parameters is out of range because the signal is noisy and incapable of being properly read by the system, the input pulse signal would not be correctly sorted. This type of incorrect sorting can occur even if the remaining nine signal parameters match the corresponding preset signal parameter values exactly. Since the system is so inflexible and incapable of sorting inputs having only one corrupt input pulse parameter, optimal results for sorting real data only approach approximately 88% accuracy. The adaptive binning system is also undesirable because it can not easily provide a "joint" quality measurement of system performance and sorting accuracy.
It has been proposed by Donald F. Specht, in his article, "Probabalistic Neural Networks for Classification, Mapping, or Associative Memory", published in the Proceedings of the 1988 IEEE International Conference on Neural Networks, Vol. 1, pp. 525-32, July 1988, to use a probabilistic neural network (PNN) to recognize input signals based upon a priori test data. Specht proposed using a PNN to search incoming data signals for a priori data patterns. The a priori test data is essentially a library or directory of patterns representing a database for the system. The probabilistic neural network developed by Specht is a multi-layer feed-forward network which uses sums of Gaussian distributions to estimate a probability density function based upon a group of a priori training patterns. The estimated probability density function is then used to sort and match new input data to the a priori training patterns.
In another article, "The Use of Probabilistic Neural Networks to Improve Solution Times for Hull-To-Emitter Correlation Problems", published by the International Joint Conference on Neural Networks, Vol. 1, pp. 289-94, June 1989, P. Susie Maloney and Donald F. Specht disclose applying a probabilistic neural network to hull-to-emitter correlation problems for electronic intelligence (ELINT) systems. However, this process operates utilizing already sorted pulse data and does not use a probabilistic neural network for real time, non a priori pulse sorting. Real time, non a priori pulse sorting is difficult because real data input signals are often noisy, inaccurate, and corrupt with additional or missing signal parameter components and information. In addition, the output probability density function for a specific signal emitter may have multiple disjoint boundaries where an individual boundary may be overlapped with another emitter probability density function. Such input signal parameters cannot be accurately approximated by an n-dimensional Gaussion distribution as proposed by Specht.